Apple Falls 7.4%, Amazon Rises 15.3% (July 2026): One AI Memory Shortage Explains Both — and Why the S&P 500 Barely Noticed

Amazon's half of the reaction kept running. It closed at $284.02 on 3 August, an all-time high that carried its market value above $3trn for the first time — then a Form 144 noticed a proposed sale of 15,000,000 shares for the account of Jeff Bezos with an aggregate market value of $4,073,700,000, the stock fell more than 2%, and the Form 4 filed two days later showed that 1,209,649 shares, about $346.5m, had actually changed hands. The session that set all of it up was 31 July: Apple closed at $308.91, down $24.52 or 7.4%, having been down close to 10% intraday — more than $400bn of market value, and enough to hand the world's-most-valuable-company title back to Nvidia days after Apple reclaimed it. Amazon closed up 15.3%. Yet the S&P 500 rose 0.7% to 7,489.72, because the two moves very nearly cancelled each other out.

The cause on both sides is the same physical shortage. AI data-centre construction has absorbed the advanced fabrication and memory capacity that consumer hardware also needs; that shows up as a cost and supply problem for the company selling phones and as accelerating revenue for the company renting servers. Apple's June quarter was excellent and its September guidance was not; Amazon's free cash flow was deeply negative and its cloud growth was accelerating. In each case the market marked the forward line and ignored the historical one.

What the settled session showed

The reactions that matter are the ones the regular session prices, and on 31 July it priced both far more forcefully than the thin extended-session quotes had.

Apple closed at $308.91, down $24.52 on the day, a fall of 7.4%. It had been down close to 10% earlier in the session before recovering into the close. The move took more than $400bn off Apple's market capitalisation and returned the title of world's most valuable company to Nvidia, days after Apple had reclaimed it. Wire copy filed before the close, when the stock was near its lows, carried a projected loss of nearly $500bn — a useful illustration of why an intraday figure and a settled one are different numbers (Reuters).

Amazon closed up 15.3%, roughly double the extended-session move, and the strongest single-session response of the reporting week.

The rest of the tape did not follow Apple down. The S&P 500 rose 52.09 points, or 0.7%, to 7,489.72; the Dow Jones Industrial Average added 276.97 to 52,485.03; and the Nasdaq composite climbed 251.68, or 1.0%, to 25,373.85. Brent crude rose 1.2% to $87.93 and the US 10-year Treasury yield edged up to 4.71% from 4.68%, which is where the day's inflation anxiety actually showed up.

What actually happened

Apple reported fiscal third-quarter revenue of $109.42bn, up 16% from $94.0bn a year earlier and above a consensus figure near $108.65bn, with diluted earnings per share of $2.02, up 29% from $1.57 and ahead of roughly $1.89 expected. Net income was $29.8bn against $23.4bn. Gross margin came in at 50.1% versus 46.5% a year earlier, including approximately 2 percentage points of favourable impact from tariff refunds, which also added about $0.11 to diluted EPS. Tim Cook described it as "our strongest June quarter ever, with double-digit revenue growth across iPhone, Mac and Services" (Apple newsroom).

The segment detail was strong almost throughout: iPhone $54.3bn, up 22%; Mac $10.4bn, up 29%; Services $30.7bn, up 12%; Wearables, Home and Accessories $7.9bn, up 6%; and Greater China up 22.4% to a June-quarter record. Only iPad fell, down 6% to $6.2bn. Across the first nine months of fiscal 2026 Apple generated roughly $117bn of operating cash flow against $81.8bn in the same period a year earlier.

Amazon's second quarter brought net sales of $200.6bn, up 20% from $167.7bn and above both consensus near $196.5bn and the top of its own $194-199bn guidance. Operating income rose 43% to $27.5bn, above the top of its guided $20-24bn range. AWS net sales rose 37% to $42.2bn, reaching a $169bn annualised run rate, with AWS operating income up 64% to $16.6bn — a 39.4% margin. North America revenue rose 16% to $116.2bn with operating income of $9.1bn; International rose 15% to $42.2bn with operating income of $1.7bn. Andy Jassy's stated summary was direct: "AWS is booming, growing 36.7% year-over-year in Q2—our fastest growth in 18 quarters" (company results release).

Why Apple beat everything and still fell

Apple cleared every bar set for the quarter it reported. It beat consensus revenue and EPS, came in above the top of its own gross-margin guidance, set June-quarter records in its three largest revenue lines, and grew Greater China 22.4% — the region that had been the most-watched risk. On the reported quarter alone the print was close to unambiguously strong.

Then the guidance arrived. Apple guided September-quarter revenue growth to 9-11% year-over-year, against the 16% it had just delivered and against a Wall Street expectation of roughly 12% — so the range missed the street at its top end, not merely at its midpoint. Gross margin was guided to 47-48% including approximately 1 percentage point of tariff benefit. Two specific causes were named: a 2.5-point foreign-exchange headwind, and supply constraints on advanced semiconductor nodes that management expected to increase significantly during the September quarter, affecting iPhone, iPad and Mac.

The mechanism worth extracting is structural rather than specific to Apple. An earnings release does two things at once: it resolves a quarter the market has already largely modelled, and it issues new information about a quarter the market has not. Only the second is genuinely new. When the two point in the same direction the reaction is easy to read; when they conflict, the forward line normally wins, because a share price discounts future cash flows and the reported quarter has already become a known quantity by the time it is announced. This is why an initial move on the release headline so often reverses once the call begins, and it is the single most common way a "beat" produces a fall.

Why Amazon's negative free cash flow was rewarded

The evening before, Meta had fallen after its free cash flow dropped to $784m from $8.55bn while costs grew roughly twice as fast as revenue. A naive reading of that reaction would be that the market had turned against cash-flow deterioration in the AI buildout. Amazon's print refutes it directly.

Amazon's free cash flow over the trailing twelve months was negative $7.6bn, against positive $18.2bn a year earlier. Purchases of property and equipment reached $173.0bn on the same trailing basis, up $65.4bn year-over-year, which the company attributed primarily to AI investment. By any cash-flow screen this is a more extreme position than Meta's. The shares rose about 8%.

What separated them was the same thing that separated Microsoft from Meta: whether the revenue attached to the spending was speeding up. AWS growth went from 28% in the first quarter to 37% in the second, its fastest in 18 quarters, against a consensus nearer 31%. AWS operating margin came in at 39.4% against a consensus around 33.8%. Both are direct evidence that committed capacity is converting into billed revenue faster than the depreciation it creates — which is the only condition under which a capital-heavy strategy can carry negative cash flow comfortably. The settled close of +15.3%, nearly double the after-hours move, says the market grew more convinced of that reading rather than less once it had a full session to think. It is the same verdict it had delivered to Microsoft two days earlier.

So the market is not pricing capital expenditure as good or bad, and it is not pricing free cash flow as a level. It is pricing the lag between money committed and revenue arriving, and it charges a company for a widening one. Amazon's lag narrowed while it spent more than ever; Meta's widened. Both facts fit one rule.

The memory shock arrived — through a different line than expected

Before either report, the shared risk identified here was the memory cost shock: DRAM and solid-state prices had risen steeply through 2026 as AI infrastructure competed for fabrication capacity, with Gartner estimating a combined increase near 130% by the end of 2026. The expectation was that it would reach Apple through the device bill of materials and gross margin, and Amazon through server costs, depreciation and AWS margin.

It arrived — and in neither company's reported margin. Apple's underlying gross margin held inside its guided band at about 48.1%. Amazon's AWS margin went up, to 39.4% from a consensus near 33.8%. Anyone using reported margins as the detector would have concluded the shock had not materialised.

It had. On Apple's side it surfaced in two places instead. First, in forward guidance: CFO Kevan Parekh attributed more than 100% of the guided sequential margin decline to memory cost increases, with carry-in inventory and cheaper non-memory components providing partial offsets that are expected to shrink. Second, in pricing — Apple raised Mac and iPad prices in June. Tim Cook's stated explanation was unusually blunt for the company: Apple had "reluctantly raised prices" because "we're in what I would characterize as a 100-year flood on the memory pricing, with exponential increases in memory prices" (Tom's Guide). The iPad line, the one segment that fell, was also one of the two that had just been repriced.

The settled reaction on 31 July suggests the market read the second half of that message — availability rather than price — as the more serious one. Cook described the constraints as "very significant" and said Apple had limited options to address them, adding that a lack of flexibility in the supply chain was making it harder to keep up with demand, and that the company was also constrained on processors. Apple had been drawing on stockpiled inventory to absorb rising memory costs, and that buffer is thinning. A cost problem compresses margin, which is a number a company can guide to and a market can model. A volume constraint caps the revenue line itself, and no amount of demand fixes it. The remarks came on Cook's final earnings call as chief executive: Apple announced in April that John Ternus becomes CEO, with Cook moving to executive chairman, and Ternus takes over on 1 September.

There is a read-across the market took seriously. Apple is generally considered the most capable hardware buyer in the industry, with the scale and the supplier relationships to secure allocation ahead of almost anyone. Ben Bajarin of Creative Strategies made the obvious inference: if a firm at Apple's scale has run out of supply-chain flexibility, smaller device makers have none. That is why the AI capacity story now sets terms for companies that build no AI infrastructure at all — the same competition for fabrication capacity runs through the chip makers supplying it.

On Amazon's side the absence of an effect is itself the mechanism, not a contradiction of it. Server memory is a capital purchase: it is capitalised on the balance sheet and reaches the income statement as depreciation spread over the asset's life. A cost shock in 2026 therefore lands on AWS margin gradually across several years rather than in the quarter the hardware is bought. That lag is why the same input cost can be visibly reshaping one company's guidance while being invisible in the other's margin during the same three months.

The number to distrust, confirmed

Before the print, this post argued that Amazon's EPS consensus was not a usable figure, because published estimates ranged from roughly $1.82 to $2.26 depending on the compiler — a spread wider than a typical surprise. The result settled the question emphatically.

Amazon reported net income of $62.6bn, or $5.75 per diluted share, against $18.2bn and $1.68 a year earlier. Inside that sat $53.4bn of non-operating pre-tax other income, which the company attributed primarily to its investments in Anthropic. The reported EPS therefore exceeded the top of the consensus range by roughly 150%, and the excess measures a change in the carrying value of a stake rather than anything about selling cloud capacity or retail goods.

This is why the guided lines are the honest scoreboard: net sales against $194-199bn and operating income against $20-24bn, both of which Amazon exceeded, and both of which describe the operating business. The general rule holds well beyond this company. Where a firm issues explicit guidance, that range is the cleanest comparison available, because it is the one where both sides use the same definition. Consensus is an average of independently constructed models; guidance is a single stated basis. It is the same reasoning behind preferring a central bank's own published projection to a survey of forecasters when judging whether a decision was a surprise.

Where it left the index

Index concentration is arithmetic, and the arithmetic resolved almost too neatly to be believed. Apple carried roughly 7.4% of the S&P 500 in late July 2026 and Amazon roughly 3.6%, so a move in each transmits that share of itself into the index before any other stock reacts. Apple fell 7.4%; Amazon rose 15.3%.

The two largest reporters of the week, carrying about 11% of the index between them, contributed approximately nothing to it on net. The 0.7% the S&P 500 actually gained came from the rest of the market. (Treat the weights as approximate and the contributions as first-order: published weights drift daily, and the official calculation runs through a divisor and float adjustments rather than a simple weighted average. The scale of the offset is the point, not the third decimal.)

That is worth holding onto, because "two mega-caps report on the same evening" is routinely described as a large index risk. It is only a large directional risk when the results point the same way. When they diverge, the concentration that would have amplified a shared surprise cancels much of it instead — while still leaving the sector-level information intact underneath. The clearest demonstration is that the Nasdaq composite, the more technology-heavy of the two indices, rose 1.0% on the day Apple had one of its worst sessions since the March 2020 crash. Concentration does not mean a large stock drags the index wherever it goes; it means a large stock gets a large vote, which the rest of the market can still outvote.

The regular session on 30 July, which closed before either release, had already been strong on its own: the S&P 500 finished at 7,437.63, up 1.66%, and the Nasdaq composite at 25,122.17, up 2.78% (Bloomberg).

The more durable consequence sits outside both companies. Because capital expenditure by the large cloud providers is revenue for the semiconductor and power-equipment supply chain, Amazon's trailing $173bn of property and equipment purchases — up $65.4bn — is a demand signal for firms that reported nothing that evening. And it connects back to the memory story: the same AI capacity demand that lifts that spending is what repriced the memory Apple buys.

Does it reach the currencies? Barely

Because this site scores currencies, it is worth being straight about how weakly this reaches them. Mega-cap earnings touch the eight majors through risk sentiment — one of the five factors the meter scores — not through the rates channel that drives most large FX moves. That relationship is weak and unstable: the long-run correlation between US equities and the dollar is only around -0.26 and flips sign for months at a time, as set out in the equity-FX feedback loop. An evening where the two largest reporters moved in opposite directions is a weaker dollar signal still. The live factor read is on the USD currency page.

Where the AI capital cycle does become a macro variable is credit rather than FX. The Bank of England devoted part of its July 2026 Financial Stability Report to AI-related systemic risk, including the growth of debt financing by AI companies — covered in our note on that report. Amazon's sale of $25bn in bonds to help fund its data-centre buildout, alongside trailing free cash flow of −$7.6bn, is a concrete instance of the trend that report flagged.

What happened next: $3trn on Monday, a $4.07bn notice on Tuesday

Amazon's reaction did not stop at the 31 July close. It closed up about 4% at $284.02 on Monday 3 August, an all-time high and its best single session since 5 May, which carried its market capitalisation above $3trn for the first time (CNBC) and made it the fifth company ever to trade at that size (Bloomberg). On the same call, CEO Andy Jassy had raised the 2026 capital expenditure figure to about $220bn from the $200bn projected in February, and said that even at that level Amazon would "still not have enough capacity to meet all the demand we have in 2026" — the same shortage that cost Apple margin, seen from the side that rents the capacity out.

Then, on the day the record printed, a Form 144 was filed for the account of Jeff Bezos. It notices 15,000,000 common shares with an aggregate market value of $4,073,700,000, to be sold on Nasdaq through Morgan Stanley Smith Barney LLC Executive Financial Services, with an approximate sale date of 3 August 2026; the shares were acquired as founder stock in July 1994, and the sale is made under a Rule 10b5-1 trading plan adopted on 14 November 2025. The filing also records that 220,200 shares were contributed to non-profit organisations on 4 May 2026 (SEC Form 144).

It was reported on Tuesday 4 August as a $4bn sale, and Amazon fell more than 2% (CNBC). The tape that day was doing the opposite: the S&P 500 gained 1.79% to a record close of 7,736.52, its first close above 7,700, the Dow added 907.47 points or 1.71% to 54,085.88, and technology rose 4% (CNBC). A stock falling 2% into a 1.8% index rally is a stock-specific reaction by definition, so the filing is the thing to examine.

A notice is not a sale, and the Form 4 settles which was which

The two documents involved answer different questions, and only the second one describes stock that reached the market.

A Form 144 is a notice of proposed sale. Under Rule 144 it must be transmitted at the same time the sell order is placed with the broker, and the person filing it must have "a bona fide intention to sell the securities referred to in the notice within a reasonable time" — an intention, over a three-month window, not an execution (17 CFR 230.144). The same rule caps what an affiliate may sell across three months at the greatest of 1% of the shares outstanding or the average weekly reported trading volume over the preceding four calendar weeks. For Amazon, 1% is about 107.9m shares, so the 15,000,000 noticed here is roughly 0.14% of the company and nowhere near the limit. The number in the notice is bounded by intent, not by the rule.

The Form 4 is the other half. Section 16 requires it "before the end of the second business day following the day on which the subject transaction has been executed" (17 CFR 240.16a-3), which is why the answer arrived on 5 August for a 3 August trade. It reports 1,209,649 shares disposed of on 3 August at a weighted-average price of $286.4083, with the highest sale at $287.00 and the lowest at $286.04 — about $346.5m, or 8.1% of the amount noticed. It leaves 879,739,004 shares held directly, some 8.2% of the company, plus 5,087,266 held indirectly through trusts and a limited liability company whose beneficial ownership the filer disclaims (SEC Form 4).

The implied prices are worth a second look, because they explain where the headline number comes from. Divide the noticed value by the noticed shares and the Form 144 is priced at $271.58 a share — below the $286.4083 the stock actually fetched that day, and well below the $284.02 close. The dollar figure in the notice is a reference price attached to a share count, and it can be stale in either direction by the time anything trades. Fifteen million shares at the realised price would have been $4.30bn, not $4.07bn.

That gives the reaction a size to compare against. A 2% fall from $284.02 is roughly $61bn of market value on that share count. The documented supply behind it, as of 6 August, is $346.5m. The market value that moved was on the order of 175 times the stock that demonstrably moved it, which rules out absorption as the mechanism — 1.2m shares is not a supply event for a company that turns over tens of millions a day. What repriced was the interpretation of a filing, and interpretations are revised when the next document lands. Any further tranche must itself appear on a Form 4 within two business days of execution, so the question is answered on a schedule rather than left open.

One thing the filings do establish is chronology. Both record the trading plan as adopted on 14 November 2025, which is more than eight months before the quarter in question was reported and nearly nine before the $3trn crossing. The purpose of a prearranged plan is to fix trade instructions in advance so that timing is not a decision taken with current information. That is what the date shows, and it is all it shows: nothing here supports any inference about anyone's view of the company, and none is offered. The pattern is also not new — the same reporting owner's filing history on EDGAR shows Form 4 disposals in July, August, September and November 2025.

The structural point generalises past this company. A concentrated index inherits the disclosure calendar of its largest members, so a document with no bearing on operations can move a percentage point of a benchmark simply because of where it is filed. Amazon carried roughly 3.6% of the S&P 500 in late July 2026, so a 2% move in it is about 0.07 percentage points of first-order index drag before anything else reacts — visible, but small enough that the index closed at a record on the same day. The identical mechanism runs through lock-up expiries, where scheduled supply is known in advance and priced imperfectly anyway: the SpaceX first-earnings note covers a 911.5m-share unlock arriving on a fixed date, which is the same variable with a calendar attached instead of a filing.

The scorecard, line by line

These were the lines identified in advance as carrying information, with what each actually resolved.

One qualification keeps this honest. A single evening identifies what the market weighted on one occasion; it does not establish a rule the market is bound to repeat. What transfers is the mechanism — forward guidance outranks the reported quarter, and capital intensity is judged on the lag between spending and revenue, not on the spending itself. The specific outcomes do not.

The week that followed added a second mechanism worth keeping, and it has nothing to do with either company's operations. When a filing rather than a result moves a mega-cap, the useful discipline is to identify which document you are reading and what that document is capable of establishing. A notice of proposed sale sets a ceiling over three months; a Form 4 records a trade within two business days. Between 3 and 5 August the difference between those two things was $4.07bn and $346.5m — the same event, described by the tape once and by the record twice.

To learn how pip theory builds its fundamental currency-strength scores, see the methodology overview.

Educational macro context only — not investment advice.